What is un­su­per­vised learning?

In un­su­per­vised learning, an AI model is trained to discover hidden patterns, re­la­tion­ships and sim­i­lar­i­ties using unlabeled data.

What is un­su­per­vised machine learning?

Un­su­per­vised learning is a data analysis method for ar­ti­fi­cial in­tel­li­gence. With this approach, an ar­ti­fi­cial neural network looks for sim­i­lar­i­ties among various input values. During un­su­per­vised learning, a computer attempts to recognize patterns and struc­tures in the input data on its own.

Un­su­per­vised learning is the opposite of su­per­vised learning, whereby de­vel­op­ers maintain complete control over the whole process and clearly define the learning outcome. With this method, the training data needs to be manually labeled or cat­e­go­rized be­fore­hand, which requires a sig­nif­i­cant in­vest­ment of time.

How does un­su­per­vised learning work?

Un­su­per­vised learning is an ar­ti­fi­cial neural network that analyzes a large amount of in­for­ma­tion, and uses that in­for­ma­tion to identify con­nec­tions, patterns and sim­i­lar­i­ties among the data. Different tech­niques are utilized to carry this out. One technique that this method uses is clus­ter­ing. With this technique, al­go­rithms group data points together based on the sim­i­lar­i­ties they share with each other.

For example, if a program is presented with pictures of cats and dogs, it will initially sort all of the dog pictures and cat pictures into distinct cat­e­gories. However, unlike with su­per­vised learning, the outcome is not pre­de­ter­mined. Un­su­per­vised machine learning al­go­rithms make these decisions on their own. Their decisions are based on the sim­i­lar­i­ties and dif­fer­ences within the pictures, for example, the color of an animal’s fur.

Another process is called as­so­ci­a­tion. With this approach, data is cat­e­go­rized based on at­trib­ut­es that it has in common with other data. The al­go­rithms’ task is to identify objects that are related. They don’t have to be identical or similar though. Using the example of the dog photos from above, an un­su­per­vised learning algorithm that is using as­so­ci­a­tion wouldn’t group all the dogs together but might associate leashes with dogs.

To display this video, third-party cookies are required. You can access and change your cookie settings here.

What is un­su­per­vised learning used for?

There are many practical examples of un­su­per­vised learning. Because it enables programs to learn game rules and strate­gies for winning, one way it can be used is for gains in the stock market. Fore­cast­ers can give a program the raw data of stock prices in order toidentify exchange ac­tiv­i­ties and predict trends.

Ar­ti­fi­cial in­tel­li­gence, and un­su­per­vised learning in par­tic­u­lar, are also widely used across many other sectors. Clus­ter­ing makes it possible to aggregate groups of people, which is of great sig­nif­i­cance for marketing, where target groups are central to the de­vel­op­ment of ad­ver­tis­ing strate­gies. Marketers can use al­go­rithms for grouping people and iden­ti­fy­ing target groups.

One sector in which un­su­per­vised learning is securely anchored is speech recog­ni­tion. Voice as­sis­tants such as Siri, Alexa and Google Assistant rely on it to work ef­fec­tive­ly. These programs learn the speech patterns of their owners, and over time, are able to un­der­stand more precise speech input, even if a device owner makes a mistake when speaking or speaks with an accent.

Many smart­phones also rely on un­su­per­vised learning to help users organize their photo galleries. Through au­tonomous and un­su­per­vised learning, the device is able to recognize the same person across multiple pictures and determine sim­i­lar­i­ties in location where photos were taken from the metadata. As such, pictures can be organized by location or by the people in the pictures.

Un­su­per­vised learning is also valuable when it comes to online chatting. Many internet users have already come across chatbots, which are now used to regulate virtual con­ver­sa­tions. Bots can also recognize insults, hate speech and racist or dis­crim­i­na­to­ry comments, and either send the offensive user a warning or remove them from the chat. Automated chats work in much the same way for customer service and online ordering ap­pli­ca­tions. Whether customers use a messenger app or SMS, the bots learn au­tonomous­ly and sometimes even un­su­per­vised.

Negative example of un­su­per­vised learning: Chatbots in social media

In 2016, Microsoft was con­front­ed with the negative effects of un­su­per­vised learning. The company’s AI “Tay” had a Twitter account and learned through its com­mu­ni­ca­tion with other users on the platform. The program was simple at first, but quickly began to use smileys and learned how to construct entire sentences. The problem with Tay was that it did not evaluate its state­ments and began to make hateful state­ments about people from different countries and feminists, and even spread con­spir­a­cy theories – all within less than 24 hours. The program was neither racially nor po­lit­i­cal­ly motivated. It simply learned from people. However, it’s unclear whether some Twitter users were poking fun at the tech­nol­o­gy and pur­pose­ful­ly fed Tay with racially and po­lit­i­cal­ly con­tro­ver­sial data.

Positive example of un­su­per­vised learning: Genetic research

In contrast, within the field of genetic research, un­su­per­vised learning has produced very positive results. Clus­ter­ing is a helpful an­a­lyt­i­cal tool to analyze genetic material. Thanks to AI and its various learning methods, medical and technical fields are coming together. This has the benefit of ac­cel­er­at­ing research tremen­dous­ly. It has been predicted that hered­i­tary diseases, such as sickle cell anemia and even hered­i­tary blindness, will be treatable and even curable in the future.

What are the ad­van­tages of un­su­per­vised learning?

Machine learning does not only stand for technical progress, but also helps to relieve the pressures of everyday life across a wide range of sectors. It’s a huge asset to our everyday lives, to the economy and to research. In contrast to other learning methods like su­per­vised and re­in­force­ment learning, de­vel­op­ers are not involved in the actual training process in un­su­per­vised learning. In addition to saving time, this approach is also able to recognize patterns that have pre­vi­ous­ly gone unnoticed. This is because un­su­per­vised machine learning gives al­go­rithms the ability to develop creative ideas.

How is it different to su­per­vised and semi-su­per­vised learning?

In addition to un­su­per­vised learning, there’s also su­per­vised learning and semi-su­per­vised learning, both of which have key dif­fer­ences that dis­tin­guish them from un­su­per­vised learning. We’ll briefly explore these dif­fer­ences below.

Un­su­per­vised learning vs. su­per­vised learning

Unlike un­su­per­vised learning, the input data and the cor­re­spond­ing outputs are known in advance in su­per­vised learning. Su­per­vised learning also has different goals. Since there is already a “correct” answer for each data point, the aim of the su­per­vised method is to train the AI to produce the answers that have already been es­tab­lished as correct.

In addition to having different goals and use cases, su­per­vised learning and un­su­per­vised learning also differ greatly in terms of ef­fi­cien­cy and trans­paren­cy. Un­su­per­vised learning only needs raw data for its training and carries out pattern recog­ni­tion on its own. However, the results are often very abstract when compared with su­per­vised learning and may need to be manually analyzed afterward. In contrast, the upfront costs of su­per­vised learning are much higher because training can only be done with data that has been labeled. The labeling of data, however, means training goals are clearly defined, and final results are generally much easier to un­der­stand.

Un­su­per­vised learning vs. semi-su­per­vised learning

In semi-su­per­vised learning, both labeled and unlabeled data are used for training. The program first uses the labeled data to create a basic model for clas­si­fi­ca­tion. Using this clas­si­fi­ca­tion model, it makes pre­dic­tions for the unlabeled data. The program is then retrained using both the original labeled data and the labels that have been generated based on its pre­dic­tions. This process can be repeated it­er­a­tive­ly to refine the model.

Like su­per­vised learning, semi-su­per­vised learning is mainly suitable for clas­si­fi­ca­tion problems. As such, it differs fun­da­men­tal­ly from un­su­per­vised learning, which is primarily used for clus­ter­ing and as­so­ci­a­tion. However, like un­su­per­vised learning, semi-su­per­vised learning has rel­a­tive­ly low upfront costs.

What other learning models are there?

In addition to the learning methods above, there’s also another learning method for AI: re­in­force­ment learning. With this method, de­vel­op­ers provide signals to influence the training of the al­go­rithms, allowing the computer to learn which decisions are correct through trial and error. For each decision, the computer receives either positive or negative feedback from the training en­vi­ron­ment. This allows ar­ti­fi­cial in­tel­li­gence to recognize patterns and develop strate­gies over time that maximize positive feedback.

For example, re­in­force­ment learning could be used to train a robot to find an object in a room, with the object being placed in a different location each time the robot searches for it. The robot would receive negative feedback for col­li­sions and wasted time. Over time, the robot would develop strate­gies to optimize its search process.